基于深度学习与深度信息的原木材积检测方法  

Log volume detection method based on deep learning and depth information

作　　者：郑积仕 张世文 杨攀 余鸿晖 李少艺 郑俊杰 周文刚 ZHENG Jishi;ZHANG Shiwen;YANG Pan;YU Honghui;LI Shaoyi;ZHENG Junjie;ZHOU Wengang(School of Transportation,Fujian University of Technology,Fuzhou 350118,China;Fujian Jinsen Forestry Co.Ltd.,Sanming 353300,China)

机构地区：[1]福建工程学院交通运输学院,福州350118 [2]福建金森林业股份有限公司,福建三明353300

出　　处：《林业工程学报》2023年第1期135-140,共6页Journal of Forestry Engineering

基　　金：福建省科技厅自然基金(2018J01619);福建金森林业股份有限公司校企合作项目(CY-H-20154);福建省林业科技项目(2021FKJO6)。

摘　　要：针对工业现场整车木材摆放深浅不一且上下分层而导致的整车木材材积自动检测困难的问题,采用ZED2双目相机获取木材端面图像以及深度图像,在改进的基于掩模区域卷积神经网络(Mask region-based convolutional neural network, Mask R-CNN)实例分割模型完成对木材端面分割的基础上,对木材轮廓进行椭圆拟合,根据木材端面中心坐标以及深度图像获取每根木材的深度信息,利用整车木材上下层深度信息相差较大的特点,解决木材上下分层的问题,同时根据双目视觉视差原理,利用相机参数以及轮廓拟合后椭圆短径端点坐标,测量出木材端面检尺径,将检尺径和检尺长代入原木材积计算公式完成原木材积的检测。实验结果表明,使用本方法计算出的整车原木材积同人工检尺数据对比,材积误差率(偶数)为-1.61%,单根木材检尺径平均绝对误差(偶数)为0.33 cm,符合国家原木检尺要求。最终将模型以及材积计算算法部署于Nvidia Jetson Xavier NX平台,并在PyQt5核心库上设计了操作界面,用户仅需拍摄完整木材端面,即可实现原木材积的计算。利用本研究所提出方法可实现木材整车检尺,解决整车木材材积自动检测困难的问题,可提高木材检尺效率,减轻检尺工作人员的工作负担。In view of the problem of difficulty in automatic detection of whole vehicle log volume caused by the different depth of vehicle log and log delamination in the industrial site, the ZED2 binocular camera was used to obtain the log end face images and depth images. On the basis of the improved Mask R-CNN(mask region based convolutional neural network) instance segmentation model to examine the log end face segmentation, the log contour is ellipse fitted. According to the central coordinates of the log end face and the depth images, the depth information of each log was obtained. Using the characteristics of the large difference in the depth information of the upper and lower layers of the whole vehicle log, the problems of the different depth of vehicle log and log delamination were addressed. At the same time, according to the principle of binocular vision parallax, the diameters of the log end faces were measured using the camera parameters and the coordinates of the ellipse short diameter endpoint after the contour fitting, the log volume can be detected by substituting the diameter and length into the log volume calculation formula. The experimental results showed that the volume error rate(even number) was-1.61% and the average absolute error(even number) of the single log diameter was 0.33 cm, which met the requirements of the national log scale. Finally, the model and volume calculation algorithm were deployed on the Nvidia Jetson Xavier NX, and the operation interface was designed on the Pyqt5 core library. Users only need to shoot the complete log end face to realize the calculation of log volume. By using the method proposed in this study, the log volume of whole vehicle could be automatically detected, the efficiency of log volume measurement was improved, and the workload of workers was reduced.

关 键 词：木材整车检尺 深度学习 Mask R-CNN 双目相机 深度信息 轮廓拟合 

分 类 号：S782[农业科学—木材科学与技术] TP391.4[农业科学—林学]